Agent memory is a policy, not a database

Retrieval is a decision

An agent memory system is often drawn as storage plus search. That misses the hard part: every remembered item competes for scarce attention.

A useful memory policy answers three questions: what should be written, when should it be retrieved, and when should it be forgotten?

Working hypothesis

Memory should preserve decisions and their evidence before it preserves conversation. Recency is helpful, but causal relevance is better.

I’m testing a simple score that combines relevance, confidence, reuse, and age. The aim is not perfect recall. It is better judgment under a small context budget.